Add Code Execution

Add Code Execution to AG2 Agents

You are an expert at setting up code execution in AG2 agent workflows. When the user wants agents that can write and run code:

1. Understand the Requirements

Ask the user:

• What language should the code be in? (Python is best supported)
• Should execution be sandboxed? (Docker for production, local for development)
• What packages/libraries does the code need access to?
• Should there be a human approval step before execution?

2. Two-Agent Code Execution (Recommended)

import os
from autogen import ConversableAgent, LLMConfig
from autogen.coding import LocalCommandLineCodeExecutor

llm_config = LLMConfig(
    {"model": "gpt-4o-mini", "api_key": os.environ["OPENAI_API_KEY"]}
)

# Coder writes code
coder = ConversableAgent(
    name="coder",
    system_message="""You are an expert Python developer.
Write code to solve the user's task. Always put code in ```python blocks.
When the task is complete, reply with TERMINATE.""",
    llm_config=llm_config,
    human_input_mode="NEVER",
)

# Executor runs code and returns output
executor = ConversableAgent(
    name="executor",
    llm_config=False,
    human_input_mode="NEVER",
    code_execution_config={
        "executor": LocalCommandLineCodeExecutor(
            work_dir="./coding_output",
            timeout=60,
        ),
    },
    is_termination_msg=lambda msg: "TERMINATE" in msg.get("content", ""),
    max_consecutive_auto_reply=10,
)

result = await executor.a_run(
    coder,
    message="Create a script that fetches the top 10 Python packages from PyPI and saves them to a CSV.",
)
await result.process()

3. Docker Execution (Production)

from autogen.coding import DockerCommandLineCodeExecutor

executor = ConversableAgent(
    name="executor",
    llm_config=False,
    human_input_mode="NEVER",
    code_execution_config={
        "executor": DockerCommandLineCodeExecutor(
            image="python:3.11-slim",
            work_dir="./coding_output",
            timeout=120,
        ),
    },
)

4. Human-Approved Execution

For workflows where a human should review code before running:

executor = ConversableAgent(
    name="executor",
    llm_config=False,
    human_input_mode="ALWAYS",  # Asks for approval before each execution
    code_execution_config={
        "executor": LocalCommandLineCodeExecutor(work_dir="./output"),
    },
)

5. Code Execution in Group Chat

from autogen.agentchat import run_group_chat
from autogen.agentchat.group.patterns import RoundRobinPattern

planner = ConversableAgent(
    name="planner",
    system_message="You break tasks into coding steps.",
    llm_config=llm_config,
    human_input_mode="NEVER",
    description="Plans the coding approach.",
)

coder = ConversableAgent(
    name="coder",
    system_message="You write Python code based on the plan. Put code in ```python blocks.",
    llm_config=llm_config,
    human_input_mode="NEVER",
    description="Writes Python code.",
)

executor = ConversableAgent(
    name="executor",
    llm_config=False,
    human_input_mode="NEVER",
    code_execution_config={
        "executor": LocalCommandLineCodeExecutor(work_dir="./output", timeout=60),
    },
    description="Runs Python code and returns results.",
)

user = ConversableAgent(name="user", llm_config=False, human_input_mode="NEVER")

result = run_group_chat(
    pattern=RoundRobinPattern(
        initial_agent=planner,
        agents=[planner, coder, executor],
        user_agent=user,
    ),
    messages="Analyze the iris dataset and create a classification model.",
    max_rounds=12,
)

6. Rules

• Separate coder (LLM) from executor (no LLM) — don't give one agent both roles
• Use DockerCommandLineCodeExecutor for untrusted code or production
• Always set timeout to prevent infinite execution
• Always set max_consecutive_auto_reply to limit retry loops
• Set work_dir to isolate output files
• Use human_input_mode="ALWAYS" during development for safety
• The executor extracts code from ```python blocks automatically
• Capital L in LocalCommandLineCodeExecutor (not Commandline)